Supermicro liquid and air-cooled systems will be validated and sold by Cisco from October, in a reference architecture built for neocloud and sovereign cloud buyers rather than conventional enterprise ones.
Cisco has brought Supermicro into its Secure AI Factory with NVIDIA, adding third-party liquid and air-cooled rack-scale servers to an architecture the company had until now assembled largely around its own compute portfolio. The Supermicro systems will be validated, sold and supported by Cisco as part of its wider AI infrastructure range, with availability beginning in October 2026.
The addition targets the density problem that has come with the current generation of accelerators. Racks built for trillion-parameter training and high-throughput inference draw far more power than conventional enterprise kit, and cooling has become the practical limit on how much compute an operator can put in a hall. Cisco is now offering rack-to-fabric liquid cooling, pairing its liquid-cooled networking systems with Supermicro servers, on platforms including NVIDIA Vera Rubin NVL72 and NVIDIA HGX Rubin NVL8.
“We are at the beginning of one of the largest datacenter buildouts in history,” said Jeetu Patel, President and Chief Product Officer at Cisco. “Every organisation is racing to scale AI, but speed only counts if it comes with control of data, managed token costs, and real ROI. It starts with the right infrastructure: compute and networking, delivered as an integrated solution that’s easy to deploy and secure from day one.”
Much of the expansion is aimed at neoclouds and sovereign cloud operators rather than at conventional enterprise buyers, and the reference architecture has been built to comply with the NVIDIA Cloud Partner programme.
Cisco Silicon One switches handle the front-end network, switches based on NVIDIA Spectrum-X handle the back-end, and both are managed through Cisco Nexus One. Cisco says it is the only NVIDIA technology partner using its own switches and network operating system inside an NCP compliant solution, a claim the company has not accompanied with a published comparison.
Justin Boitano, vice president for Enterprise AI at NVIDIA, said AI factories are revenue-generating infrastructure, “where compute produces intelligence, and intelligence drives revenue”. He said the expanded architecture would help enterprises and neoclouds reach production faster.
Cisco is also introducing Cisco Validated Infrastructure Services, aligned to NVIDIA Infrastructure Services, to certify that deployed racks match the reference design, and is funding a large-scale AI lab to develop testing tools for it. Operations sit with NVIDIA AI Enterprise software and Cisco Cloud Control, which the company says will correlate job health with compute, network interface, optics and network performance data.
No customer names, pricing, deployment timelines or performance baselines were disclosed, so the claimed gains in supply certainty and time to production cannot be measured against Cisco’s existing stack.
The more telling detail is the one Cisco does not spell out: a company that has spent two years selling AI infrastructure as its own end-to-end system has now made room inside that system for someone else’s servers.
